江正宇 Email & Phone Number

江正宇 is a 國立交通大學學生 at Accton.

• 

  Intern

  Accton

  May 2024 - Present - Hsinchu City, Taiwan, Taiwan

  Focused on developing an offline LLM model, with the goal of building an SQL expert model capable of automatically generating SQL queries and retrieving corresponding data from a database, then responding to users with text and charts. Due to hardware limitations, we conducted research using small models with 7B to 13B parameters.We utilized tools like Ollama and Open WebUI to provide a more user-friendly interface, instead of relying solely on Linux command-line operations. Through Ollama’s customization features, we adjusted model parameters, temperature, and prompt design, steering the model towards becoming a specialized SQL expert.We introduced Retrieval-Augmented Generation (RAG) technology, using the Vanna package to connect VectorDB and the database, improving search efficiency and accuracy. Fine-tuning of the model began using a SQL QA dataset generated by GenAI, continuously optimizing the model’s SQL generation capabilities.Although significant progress was made throughout the process, due to the lack of specific evaluation metrics, we are currently unable to quantify the model's performance improvements.專注於開發離線LLM模型，目標是建立專精於SQL的生成模型，能自動生成SQL查詢並從資料庫中提取對應數據，並以文字和圖表回應使用者的問題。由於硬體資源不足，使用7B~13B參數的小模型進行研究。使用了Ollama和Open WebUI等工具，為系統提供了更友好的使用者介面，而非單純的Linux命令行操作。透過Ollama的自定義功能，調整模型的參數、溫度和prompt設計，使模型更傾向於成為專業的SQL專家。導入RAG技術，使用Vanna package實現了VectorDB與資料庫的連接和搜尋功能，提升了檢索的效率和準確性。開始進行模型的Fine-Tuning，使用GenAI生成的SQL QA資料集進行訓練，不斷優化模型的SQL生成能力。雖然過程中取得了顯著的進步，但由於缺乏具體的模型評估指標，目前無法量化模型的效能提升。